False news detection method based on large language model cognitive enhancement

Abstract

The invention particularly relates to a false news detection method based on large language model cognition enhancement, which comprises the following steps of: constructing a user portrait model and a personality characteristic model by a cognition-based comment generation module, interacting the user portrait model and the personality characteristic model with a large language model, and generating a personification comment conforming to a user cognition mode; the multi-view comment learning module is used for extracting semantic, emotion and style characteristics from the generated anthropomorphic comments, analyzing the consistency and inconsistency of the comments and news contents based on a large language model, and generating a comment summary; the expert intelligent automatic screening module is used for guiding a large language model to generate expert opinions matched with the news samples through an experience matching strategy and a domain enhancement strategy; and the interactive fusion and classification module fuses the news content, the comment summary and the expert opinions, and outputs the news authenticity probability through a cross attention mechanism and a classifier. Through the design of a plurality of modules, the accuracy and robustness of false news detection are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a fake news detection method based on large language model cognitive enhancement. Background Art

[0002] In related technologies, the widespread spread of fake news has seriously disrupted the order of Internet information dissemination, and has had a profound negative impact on social development, public health, economic prosperity, and even political security, and has become a major threat to network content security. With the advent of the era of large language models (LLMs), fake news detection based on large language models has become a core issue that needs to be solved urgently.

[0003] Existing fake news detection techniques have gone through three stages of development. The first stage is the feature engineering stage, which manually collects simple explicit features (such as content length, number of likes, etc.) in social media to identify fake news. Although this method can discover some effective features, it has an obvious disadvantage, that is, it requires a lot of manpower. With the rise of deep learning, fake news detection has entered the second stage based on deep learning. This stage mainly uses deep learning technology to learn the semantic, emotional and position characteristics of news content, while capturing comment-related, dissemination-related and user portrait-related features in metadata (such as social context). Although this type of method is sophisticated in design and has significantly improved performance, its complex network structure and a large number of parameter tuning processes are still important challenges. The third stage focuses on fake news detection based on large language models. Large language models have made significant progress in many application fields, and there are currently many methods that have introduced them into fake news detection tasks.

[0004] Large language model-based methods can be mainly divided into two categories: One is to directly input unverified news into the large language model and design simple prompting strategies to make the model output the authenticity of the news. However, research shows that the performance of this method is usually difficult to rival that of deep learning-based techniques. The other method utilizes the generative ability of large language models (LLMs) to construct user-generated simulators, generate diverse comments for specific news, and use them as semantic supplements to enhance the ability to detect fake news. This method has great potential, especially in generating comments and designing prompting strategies, and can provide better interpretability. However, these methods are still in the preliminary experimental stage and have some significant drawbacks: 1) The granularity of user cognitive modeling is relatively rough, and the generated comments are difficult to make full use of: Existing methods usually construct user portraits based on simple attributes (such as gender, age, etc.) to model user cognition, generate prompts through these portraits, and let the large language model generate news comments. This method lacks a fine-grained description of user cognition, resulting in the generated comments being difficult to truly reflect user cognition. In addition, after generating multiple comments, semantic clustering is usually only focused on for detection. This approach ignores the multi-perspective features (such as sentiment or stance features) and inconsistent semantics (such as skeptical voices) in the comments, thus limiting the extraction of crucial credibility information in fake news detection. The understanding of user cognition and the generated content is still imperfect and overly simplified, affecting the overall detection effect. 2) The method based on fixed prompt content lacks pertinence in analyzing news authenticity and is prone to generating noisy data: Existing methods often rely on manually designed expert modules to guide LLMs to evaluate news authenticity from specific perspectives (such as biased or inflammatory content) through predefined prompts. This method is based on the entire dataset rather than specific news instances, resulting in low specificity and effectiveness of the prompting strategy. In addition, the manually designed prompts may introduce noisy data, further weakening the model performance.

[0005] Therefore, designing a large language model-based fake news detection method with strong pertinence, detailed user modeling, and reliable performance has become a key research issue in the field of fake news research.

[0006] It should be noted that the information disclosed in the above background technology section is only used to strengthen the understanding of the background of the present invention, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0007] The present invention provides a fake news detection method based on large language model cognitive enhancement, which can achieve fake news detection with strong pertinence, detailed user modeling, and reliable performance, and can thus overcome the defects existing in the prior art to a certain extent.

[0008] Other features and advantages of the present invention will become apparent from the following detailed description or will be partly learned through the practice of the present invention.

[0009] According to a first aspect of the present invention, there is provided a method for detecting false news based on cognitive enhancement of large language models, and a detection model integrating an audience cognitive enhancement and an expert intelligence automatic screening mechanism is designed. The model includes:

[0010] A cognition-based comment generation module constructs a user portrait model and a personality trait model, interacts the user portrait model and the personality trait model with a large language model, and generates anthropomorphic comments that conform to the user's cognitive pattern.

[0011] A multi-perspective comment learning module extracts semantic, emotional, and style features from the generated anthropomorphic comments, analyzes the consistency and inconsistency between the comments and the news content based on the large language model, and generates a comment summary.

[0012] An expert intelligence automatic screening module guides the large language model to generate expert opinions adapted to the news samples through an experience matching strategy and a domain reinforcement strategy.

[0013] An interaction fusion and classification module fuses the news content, the comment summary, and the expert opinions, and outputs the news authenticity probability through a cross-attention mechanism and a classifier.

[0014] In some exemplary embodiments, the user portrait model is based on six key user-related categories, including age, gender, education level, marital status, geographical location, and media preference, and generates comments reflecting different user categories through interaction with the large language model.

[0015] In some exemplary embodiments, the personality trait model is based on the five major personality traits, including openness, conscientiousness, extraversion, agreeableness, and neuroticism, and generates comments reflecting different personality traits through interaction with the large language model.

[0016] In some exemplary embodiments, the cognition-based comment generation module further includes a cognitive focus mechanism:

[0017] Screen the top k1 comments with the highest relevance to the news content through semantic similarity.

[0018] Randomly select k2 remaining comments, combine them and input them into the large language model to generate a new round of comments.

[0019] In some exemplary embodiments, the multi-perspective comment learning module distinguishes comments through the following steps:

[0020] Screen consistent comments based on cosine similarity.

[0021] Screen inconsistent comments based on KL divergence;

[0022] Summarize consistent and inconsistent comments from the dimensions of semantics, sentiment, and style.

[0023] In some exemplary embodiments, the experience matching strategy includes:

[0024] Calculate the semantic similarity between the target news and the training samples;

[0025] Select the reference sample with the highest similarity to construct a prompt to guide the large language model to generate expert opinions.

[0026] In some exemplary embodiments, the experience matching strategy includes:

[0027] Extract news keywords by the 5W1H method and cluster them;

[0028] Based on the Euclidean distance, screen domain-related positive and negative samples, construct domain-specific prompts, and guide the large language model to generate expert opinions.

[0029] In some exemplary embodiments, the interaction fusion and classification module uses BERT to encode news, comment summaries, and expert opinions, realizes feature fusion through the cross-attention mechanism, and uses the softmax function to output the probability of fake news.

[0030] According to the second aspect of the present invention, there is provided a storage medium on which a computer program is stored, and when the computer program is executed by a processor, it implements the method for detecting fake news based on large language model cognitive enhancement described in the first aspect above.

[0031] According to the third aspect of the present invention, there is provided a computer program product on which a computer program is stored, and when the computer program is executed by a processor, it implements the method for detecting fake news based on large language model cognitive enhancement described in the first aspect above.

[0032] According to the fourth aspect of the present invention, there is provided an electronic device, including:

[0033] A processor; and

[0034] A memory for storing executable instructions of the processor;

[0035] Wherein, the processor is configured to implement the method for detecting fake news based on large language model cognitive enhancement described in the first aspect above when executing the executable instructions.

[0036] The false news detection method based on large language model cognitive enhancement provided by the embodiments of the present invention proposes a new detection model that integrates audience cognitive enhancement and expert intelligent automatic screening mechanism. Through the design of multiple modules, the accuracy and robustness of false news detection are improved.

[0037] First, the present invention designs a cognition-based comment generation module, which combines external identity attributes (such as gender, age) and internal personality characteristics (such as the Big Five personality traits) to simulate multi-perspective and real human comments, providing diverse evidence and insights for false news detection. Then, a multi-perspective comment learning module is proposed, which focuses on extracting multi-dimensional cognitive information from the generated comments, including semantics, stance, and style, and obtaining credibility evidence by identifying consistencies and inconsistencies.

[0038] In addition, the expert intelligent automatic screening module of the present invention provides rich prior knowledge for the model through a prompt-based strategy, combining experience matching and domain-specific enhancement methods, to assist in the directional analysis of news authenticity. Finally, the interactive fusion and classifier module strengthens the integration of comment features and news content, effectively improving the detection effect.

[0039] Compared with the prior art, the present invention has the following innovations:

[0040] 1. The present invention constructs a refined user cognitive modeling framework based on LLMs, comprehensively considering the refined user profile and personality characteristics, and can generate high-quality news comments for different user cognitions.

[0041] 2. The present invention designs an experience matching and domain reinforcement strategy, enabling LLMs to generate customized prompts for each news sample, reducing manual intervention. This method adaptively reveals the unique semantics of each article, ensures logical consistency, and effectively guides LLMs to generate accurate and insightful expert opinions.

[0042] 3. The present invention designs a unified false news detection framework, organically integrating LLMs and deep learning models, and making full use of the comments generated by LLMs to learn consistency and inconsistency features from multiple perspectives, thereby improving the detection effect.

[0043] The present invention has conducted extensive experiments on two competitive false news detection datasets, verifying the effectiveness and superiority of the present invention.

[0044] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. Brief Description of the Drawings

[0045] The accompanying drawings herein are incorporated into and constitute a part of this specification, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0046] Figure 1 It is an architecture diagram of the method of the present invention. Detailed implementation manners

[0047] Example embodiments will now be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be more complete and comprehensive, and will fully convey the concept of the example embodiments to those skilled in the art. The features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments.

[0048] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0049] In view of the deficiencies in existing false news detection methods, in this example embodiment, a false news detection method based on enhanced cognition of large language models is provided. Given a news input, the model relies on large language models (LLMs) to refine user cognition modeling, generates personalized comments and expert opinions for specific news through adaptive prompts, and learns the consistency and inconsistency features in the generated comments and opinions to enhance false news detection.

[0050] The present invention proposes a novel detection model that integrates audience cognitive enhancement and an expert intelligent automatic screening mechanism. The model is divided into the following four modules: a cognition-based comment generation module, a multi-perspective comment learning module, an expert intelligent automatic screening module, and an interaction fusion and classification module. Among them, the cognition-based comment generation module focuses on constructing a detailed user profile and modeling user personality to accurately depict user cognition, and designs different prompting strategies to promote the generation of high-quality anthropomorphic user comments by LLMs; the multi-perspective comment learning module makes full use of the generated comment information. Different from traditional methods that usually only consider capturing the semantic perspective of comments, the present invention designs a multi-perspective comment learning module to mine the consistency and inconsistency from multiple perspectives such as stance, emotion, and semantics in the generated comments to discover supporting and skeptical voices in the comments; the expert intelligent automatic screening module regards the large language model as a consulting expert, focuses on automatically docking customized prompting strategies with different news samples, and adaptively reveals the unique semantics of each article. This module ensures logical consistency by emphasizing experience matching and domain reinforcement, and effectively guides LLMs to generate accurate and insightful expert opinions; the interaction fusion and classification module integrates the relevance of the viewpoints of comments and expert opinions with the news content itself to reveal the credibility of the news.

[0051] Reference Figure 1 As shown, a detection model that integrates audience cognitive enhancement and an expert intelligent automatic screening mechanism is designed. The model includes:

[0052] A cognition-based comment generation module constructs a user profile model and a personality feature model, interacts the user profile model and the personality feature model with the large language model, and generates anthropomorphic comments that conform to the user's cognitive pattern;

[0053] A multi-perspective comment learning module extracts semantic, emotional, and style features from the generated anthropomorphic comments, analyzes the consistency and inconsistency between the comments and the news content based on the large language model, and generates a comment summary;

[0054] An expert intelligent automatic screening module guides the large language model to generate expert opinions adapted to the news samples through an experience matching strategy and a domain reinforcement strategy;

[0055] An interaction fusion and classification module integrates the news content, the comment summary, and the expert opinions, and outputs the news authenticity probability through a cross-attention mechanism and a classifier.

[0056] The above-mentioned cognition-based comment generation module is specifically described as follows:

[0057] Since news comments usually reflect the cognitive perspective of individual audiences, using large language models to generate user comments and accurately simulate user cognition is the key to docking with real-world behavior. To this end, the present invention develops a fine-grained cognition-based comment generation module, which enhances the modeling of user portraits (such as age, gender, education) and personality traits (such as openness, conscientiousness, and extroversion) to deeply simulate user cognition and ensure that the generated comments are consistent with the user's intrinsic cognitive patterns. In addition, users' comments on specific news articles are usually influenced not only by the content of the article, but also by previous comments. To take this into account, the present invention proposes a cognitive focusing mechanism that can identify and select relevant previous comments that may affect user views. By integrating these comments with the original news content and inputting the synthesized information into LLMs, this two-layer approach ensures that the generated comments more accurately simulate the behavior and cognitive tendencies of real users.

[0058] Specifically, user portrait modeling is described as follows:

[0059] Different users often have different views on the same news article. To capture these individual differences, this paper adopts six key user-related categories to build user profiles and uses large language models to generate news comments that can reflect diverse perspectives. By modeling these categories, the model can objectively capture the wide range of audiences that news may encounter in real-world scenarios. Based on the news content, LLMs generate comments according to a predefined set of prompts, thereby enhancing the diversity and granularity of comments.

[0060] For example, the categories, attributes and classification basis of user portraits are shown in Table 1:

[0061] Table 1 User portrait categories, attributes and classification basis

[0062]

[0063] For example, the user portrait modeling is constructed as follows:

[0064] User modeling: Assume that you are a Weibo user. Your gender is [gender], age is [age], hometown is [hometown location], and marital status is [marital]. You usually get news through [media preference]. Now you will be provided with an article [news content] of questionable authenticity. Please write a comment of no more than 60 words based on your identity information, which can reflect the style of a typical Weibo user.

[0065] Specifically, personality trait modeling is described as follows:

[0066] Individuals with different personality traits often exhibit different reactions when exposed to fake news. Therefore, this step aims to generate news article comments using large language models, simulating diverse user cognitions by reflecting different personality traits, thereby enhancing the diversity of comment content. To reasonably classify personality traits, the present invention adopts the internationally recognized "Big Five Personality Traits" as the framework for personality trait classification. By constructing prompts containing personality trait content, LLMs are guided to generate comments on the same news sample. Specifically, the "Big Five Personality Traits" include the following attributes: openness, conscientiousness, extraversion, agreeableness, and neuroticism. Through the prompts constructed in this way, the comments generated from different personality perspectives may reflect a wide range of emotional and cognitive responses to the same news content. This diversity of responses is crucial for tasks such as fake news detection, as it provides profound insights into how users perceive and interpret news under the influence of their specific personality traits, thereby improving the robustness and accuracy of the detection process.

[0067] Specifically, the cognitive focus mechanism is described as follows:

[0068] On social media platforms, users' cognitive responses are influenced by the behaviors of other users; when users browse news content and read comments, these interactions shape the way they form new comments. To comprehensively capture these user behavior characteristics, the present invention designs a cognitive focus mechanism, considering from two perspectives: semantic similarity and random selection. In the semantic similarity perspective, the present invention selects the top k1 comments that are most relevant to the news content viewpoints, because these comments have the greatest impact on users' formation of new viewpoints. The present invention uses an encoder to encode the news content into a vector e t , and encodes each comment into a vector e c , and then calculates the cosine similarity between the news content and each comment. Since some comments may have little association with the news content, resulting in low similarity scores, the present invention requires leaving the top k1 comments with the highest cosine similarity to the news content, while filtering out other comments with low cosine similarity to the text, which can automatically exclude comments that are irrelevant or weakly associated with the news content, while retaining the most valuable comments.

[0069] e c = Encoder(c i ) (1)

[0070] e t = Encoder(t) (2)

[0071]

[0072] where, e t ·e c is the dot product of two vectors, |e t|and|e c |is the modulus of the vector.

[0073] Similarly, in the real world, users' browsing and commenting behaviors are inherently unpredictable. To capture this randomness, the present invention randomly selects an additional k2 comments from the remaining unselected comment pool to simulate the diversity in user interactions. These randomly selected comments, together with the initially selected k1 comments, form a new set of comments presented to the virtual user, enabling it to generate a new round of a total of q comments (k1 + k2 = q). This method ensures the diversity of content and viewpoints, thus simulating the dynamic characteristics of news dissemination in the real world.

[0074] The above multi-perspective comment learning module is specifically described as follows:

[0075] Although news comments from different cognitive perspectives vary significantly, they generally reflect agreement or disagreement with the news content, conveying attitudes such as support, neutrality, or doubt. Previous studies mainly focused on the parts that are consistent with the news content, often overlooking the valuable insights contained in those comments that are contradictory to it. To make full use of user comments generated by large language models (LLMs), the present invention proposes a multi-perspective comment learning module. Different from traditional methods that mainly focus on semantic analysis, the present invention extends the learning perspective to include multiple dimensions such as semantics, sentiment, and style, enabling the model to extract valuable information from all comments, regardless of whether the comment information supports or opposes the news content, thus providing a more comprehensive understanding for fake news detection.

[0076] To capture the key information of comments related to fake news detection, the present invention selects two representative metrics to screen the comment set. Specifically, the present invention classifies all comments into two categories: "consistent comments" based on semantic similarity and "inconsistent comments" based on divergence evaluation, with the latter calculated using KL divergence.

[0077] This method first uses an encoder to encode the comments in the comment library. Then, it calculates the cosine similarity between each comment and all other comments, and selects the top q1 comments with the highest cosine similarity as consistent comments. For inconsistent comments, it is assumed that different dimensions in the comment embedding space correspond to different perspectives. To evaluate the diversity of viewpoints in each perspective, the present invention calculates the KL divergence based on these embeddings. Specifically, for each comment c i , it calculates its KL divergence from all other comments c j . The sum of the KL divergences of comment c i from all other comments is:

[0078]

[0079] Where, and are the embedded vectors of the comments, where q represents the total number of comments corresponding to each news article. After screening by the first module, the KL divergence is selected as the metric for detecting comment inconsistency because of its effectiveness in measuring the difference in probability distributions, computational simplicity, and its wide application in similar tasks compared to alternative methods such as the Wasserstein distance.

[0080] The present invention selects the q2 comments with the highest total KL divergence to highlight the differences between different perspectives, representing "inconsistent comments". The present invention guides the LLMs to summarize the consistent and inconsistent comments from dimensions such as semantics, sentiment, and style. The summaries of these comments are used to evaluate the authenticity and credibility of the news, providing a detailed analysis that integrates supporting and opposing views.

[0081] The above-mentioned expert wisdom automatic screening module is specifically described as follows:

[0082] To overcome the shortcomings of existing methods that ignore the specific context or domain of news content and lack nuanced, context-sensitive analysis, the present invention designs an expert intelligent automatic screening module that can automatically screen appropriate prompts for different news to adaptively mine the internal semantics of the news. The module includes two core news prompt adaptive matching strategies: experience matching and domain reinforcement. These two strategies work together to construct prompts, guiding the LLMs to focus on the most relevant aspects of the news article and generate expert-level opinions, thereby improving the accuracy and efficiency of fake news detection. This innovation not only enhances the automation level but also maximizes the performance advantages of the LLMs. To enhance the autonomy of the model in the decision-making process, the method of the present invention is no longer limited to simple true / false judgments, but provides a probability assessment of authenticity for each news article, enabling the model to make adaptive judgments based on the specific characteristics of the news, thus achieving a more delicate and flexible analysis.

[0083] Specifically, the experience matching strategy is described as follows:

[0084] The experience matching module aims to provide stronger credibility signal guidance for unvalidated news through existing experience. Specifically, the module constructs a semantic similarity match between the target news and the training samples, and screens out highly relevant samples from the training samples to provide more credibility features for the LLMs based on the unvalidated news, analyzing from perspectives such as content semantics, discourse structure, and expression style. The semantic similarity match can be formalized as:

[0085]

[0086] where, and e jRepresent the embeddings of news articles in the training set and the embeddings of target samples respectively.

[0087] After obtaining the reference news provided by the experience matching module, the target news and the reference news are presented to the LLMs simultaneously. Subsequently, the LLMs are guided to evaluate the authenticity of the target news from multiple perspectives, including but not limited to aspects related to logical consistency, the presence of exaggerated elements, etc. The model then provides reasons to support its authenticity assessment. This strategy is formalized as:

[0088]

[0089] where p E represents the experience matching prompt, represents the reference news article selected according to experience matching.

[0090] This multi-faceted approach ensures that the experience matching module can capture the interaction between semantic and logical factors, enhancing the accuracy and robustness of fake news detection.

[0091] Demonstratively, the experience matching prompt template is:

[0092] You will see a target news article with unknown authenticity. Please help analyze the authenticity of the target news. To assist you in this task, I will provide two simple examples: The first example is [Example 1], classified as [News Label]; the second example is [Example 2], classified as [News Label]. You should first analyze the reasons why these two examples are classified into their respective labels, and then use this analysis to evaluate the authenticity of the target news. Your answer should follow the following format: Probability of real news; Reason.

[0093] Specifically, the domain reinforcement strategy is described as follows:

[0094] Fake news in different domains usually exhibits unique characteristics. For example, political fake news may distort facts or quote remarks out of context, using inflammatory language to mislead public opinion; while health fake news tends to exaggerate the effects of certain drugs or treatment methods, spreading through the form of authoritative reports, and sometimes even spreading unproven treatment methods. Therefore, it is particularly important to incorporate domain-specific information into the context of news articles. This is mainly achieved through news classification, which can identify the domain of the article. To improve the accuracy of fake news detection, the present invention proposes a strategy combining domain classification and target guidance. Specifically, the present invention uses LLMs to automatically classify the domain of news articles according to their content. Domain classification provides the model with clearer context information, enabling it to better understand the article content and enhance its ability to evaluate authenticity.

[0095] On this basis, the present invention designs a targeted prompting mechanism to guide LLMs to evaluate the authenticity of news based on the domain-specific features of fake news. Different from traditional generalization methods, the present invention proposes a domain reinforcement strategy that focuses on identifying the typical features of fake news within a specific domain, enabling the model to more accurately detect the domain-specific characteristics of fake news. This method allows LLMs to focus on the unique signs of fake news in each domain, improving the accuracy and reliability of their judgments. In some datasets, the news sources are relatively homogeneous, and there may be ambiguity in domain classification, making traditional domain classification methods no longer applicable. To address this issue, the present invention adopts the "5W1H" (who, what, time, place, why, how) method to extract keywords from news articles and uses a pre-trained model to represent the keywords of each article. Subsequently, the present invention uses the KNN algorithm to cluster the representations of these keywords to obtain news categories at a certain abstract level.

[0096] When providing feedback to LLMs, due to the ambiguity of news categories and their differences from traditional well-defined domain classification methods, the present invention directly calculates the Euclidean distance between the test sample and the news articles in the training set. Select the k positive samples and k negative samples with the smallest distance from the test sample, and use these reference samples to construct prompting content to guide the model to evaluate the authenticity of the candidate article. This method allows the model to utilize domain-specific prior knowledge to provide a more structured reference framework for evaluating the authenticity of unknown news samples. This method effectively solves the problem of domain classification ambiguity and enhances the model's ability to evaluate the authenticity of unknown news by combining the context information of the training set. The specific steps can be formally expressed as follows:

[0097]

[0098] where p D represents the domain reinforcement prompt, represents the i-th positive reference sample, represents the i-th negative reference sample.

[0099] The above interaction fusion and classifier are specifically described as follows:

[0100] The interaction fusion and classifier module consists of two main branches. The first branch is responsible for integrating the comprehensive summary of news content and comments, and the second branch focuses on combining the news content with the corresponding expert opinions. Together, they achieve a complementary balance between audience-driven cognition and the perspective based on expert knowledge. Specifically, the present invention uses BERT as the encoder for news content, comments, and opinions to obtain semantic representations. To facilitate comprehensive information interaction, a cross-attention mechanism is introduced to fuse the comprehensive comments or opinions with the news content, aggregating key aspects such as stance, semantics, and the content itself. The cross-attention network can be described as:

[0101]

[0102] where q′ = w q q, k′ = w k k, v′ = w v v, and d is the feature dimension.

[0103] After deeply fusing the news text through the cross-attention mechanism, a comprehensive vector is obtained, which combines the audience's attitude and the expert's perspective, as well as multi-perspective comment summaries, helps to identify the authenticity of news articles, realizes rich interaction between news and opinions, and adaptively selects useful reasons as references. For the representation e t of the news article, the comment summary e soc and the opinion e o , an average pooling operation is adopted, and its formula is as follows:

[0104] e t = Encoder(New i ) (11)

[0105] e soc = Encoder(soc) (12)

[0106] e o = Encoder(o) (13)

[0107]

[0108] where soc represents the comment summary, o represents the opinion generated by the LLM, T is the sequence length, and the average pooling operation is used to obtain a single vector representation and

[0109] After obtaining the learned content of the two branches, these features are concatenated, and the formula is as follows:

[0110]

[0111] where h clsis the fusion vector, and W represents a learnable parameter matrix. Subsequently, the vector h cls is input into the MLP to obtain h' cls .

[0112] Finally, the present invention uses the softmax function to output a probability distribution for training, and more accurately discriminates the authenticity of news by utilizing the aggregated information.

[0113]

[0114] Loss function

[0115]

[0116] where y i is the true label of the news item, is the probability that the news article is predicted to be fake news.

[0117] Next, the false news detection method of the present invention is simulated and verified on two independent datasets, Weibo21 and GossipCop, and the accuracy rates reach 80.0% and 89.0% respectively.

[0118] Table 2 Performance comparison between the present invention and the baseline model on the Weibo21 and GossipCop datasets

[0119]

[0120] Table 3 Performance comparison of ablation of different modules of the present invention on the Weibo21 and GossipCop datasets

[0121]

[0122]

[0123] It should be noted that, on the other hand, the present application also provides a storage medium, which may be included in an electronic device; or may exist alone without being assembled into the electronic device. The above storage medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device is caused to implement the method described in the following embodiments.

[0124] In one embodiment, the present application provides a computer program product, including a computer program, which when executed by a processor implements the steps in the above method embodiments.

[0125] In addition, the above-mentioned drawings are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present invention, rather than for limiting purposes. It is easy to understand that the processes shown in the above-mentioned drawings do not indicate or limit the chronological order of these processes. Additionally, it is also easy to understand that these processes can be executed, for example, synchronously or asynchronously in multiple modules.

[0126] Other embodiments of the present invention will be readily apparent to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the present invention and include known common knowledge or conventional technical means in the technical field not disclosed by the present invention. The specification and embodiments are to be considered as exemplary only, and the true scope and spirit of the present invention are pointed out by the claims.

[0127] It should be understood that the present invention is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present invention is only defined by the appended claims.

Claims

1. A method for detecting false news based on cognitive enhancement of large language models, characterized in that, A detection model that integrates an audience cognitive enhancement and an expert intelligence automatic screening mechanism is designed. The model includes: A cognition-based comment generation module that constructs a user portrait model and a personality trait model, interacts the user portrait model and the personality trait model with a large language model, and generates anthropomorphic comments that match the user's cognitive pattern. A multi-perspective comment learning module that extracts semantic, emotional, and style features from the generated anthropomorphic comments, analyzes the consistency and inconsistency between the comments and the news content based on the large language model, and generates a comment summary. An expert intelligence automatic screening module that guides the large language model to generate expert opinions adapted to the news samples through an experience matching strategy and a domain reinforcement strategy. An interaction fusion and classification module that fuses the news content, the comment summary, and the expert opinions, and outputs the probability of news authenticity through a cross-attention mechanism and a classifier.

2. The method according to claim 1, characterized in that The user portrait model is based on six key user-related categories, including age, gender, education level, marital status, geographical location, and media preference, and generates comments reflecting different user categories through interaction with the large language model.

3. The method according to claim 1, characterized in that The personality trait model is based on the five major personality traits, including openness, conscientiousness, extraversion, agreeableness, and neuroticism, and generates comments reflecting different personality traits through interaction with the large language model.

4. The method according to claim 2 or 3, characterized in that, The cognition-based comment generation module further includes a cognitive focus mechanism: Screen the top k1 comments with the highest relevance to the news content through semantic similarity. Randomly select k2 remaining comments, combine them and input them into the large language model to generate a new round of comments.

5. The method according to claim 1, wherein The multi-perspective comment learning module distinguishes comments through the following steps: Screen consistent comments based on cosine similarity. Screen inconsistent comments based on KL divergence. Summarize consistent and inconsistent comments from the semantic, emotional, and style dimensions.

6. The method according to claim 1, wherein The experience matching strategy includes: Calculate the semantic similarity between the target news and the training samples. Select the reference sample with the highest similarity to construct a prompt to guide the large language model to generate expert opinions.

7. The method according to claim 1, wherein The experience matching strategy includes: Extract news keywords through the 5W1H method and cluster them. Screen domain-related positive and negative samples based on Euclidean distance, construct a domain-specific prompt, and guide the large language model to generate expert opinions.

8. The method according to claim 1, characterized in that The interaction fusion and classification module uses BERT to encode the news, the comment summary, and the expert opinions, realizes feature fusion through a cross-attention mechanism, and uses the softmax function to output the probability of fake news.

9. A storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by a processor, it implements the fake news detection method based on large language model cognitive enhancement as described in any one of claims 1 to 8.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the fake news detection method based on large language model cognitive enhancement as described in any one of claims 1 to 8.

Images